fix(macos): stop renderer injection hot loops

[ImcLiuQian]

Summary / 摘要

• Make renderer root-class application idempotent so our own class mutation no longer re-triggers the observer hot loop.
• Make early bootstrap a single, bounded lifecycle with explicit cleanup; avoid duplicate observers/timers across reconnects and navigation.
• Reject auxiliary routes such as /avatar-overlay, and close the CDP socket when domain setup fails.
• Add regression coverage for repeated injection, reconnect, navigation/renderer replacement, timeout cleanup, target filtering, and handshake failure.

Root cause / 根因

Live CDP A/B evidence confirms a Dream Skin self-triggered renderer loop: with injection enabled, the observed renderer ran ensure at about 9/s and spent about 14.1% of one core in style recalculation; after Dream Skin cleanup, style recalculation dropped to about 0.015%. An independent root/body observer saw no external class writer, while applyRootState() unconditionally rewrote the watched root class.

The historical 7.5 GiB / 200–320% CPU incident cannot be attributed solely to Dream Skin because no historical heap/socket snapshot exists. This PR fixes the confirmed Dream Skin hot loop, allocation churn, repeated bootstrap state, and handshake socket leak without claiming it was the only contributor to that high-water event.

Type / 类型

• Bug fix / 缺陷修复

Platform / 平台

• macOS

Self-check / 自测

• CODEX_DREAM_SKIN_SKIP_DOCTOR=1 /bin/bash macos/tests/run-tests.sh passed
• node macos/tests/injector-bootstrap.test.mjs — 7/7
• node macos/tests/renderer-inject.test.mjs — passed
• Live verify through installed loopback-CDP runtime
• status-dream-skin-macos.sh --json --deep: session active, injector alive, CDP OK
• Responsive smoke: 1200×900 = 2 columns; 680×900 = 1 column; no horizontal overflow

Runtime evidence / 运行时证据

A bounded 60s post-fix sample showed ensure at about 0.28/s, style recalculation about 0.32% of one core, fixed DOM/style/chrome node counts, and JS heap 374→321 MiB. Point-in-time main Renderer samples stayed far below the historical incident (roughly 6.7–16.9% CPU, 498–590 MiB RSS), but these are observations rather than a controlled long-duration baseline.

Only injector.mjs and renderer-inject.js were synced to the installed runtime for verification. Existing local CSS/theme customization was preserved; Codex itself was not restarted.

Security / 安全

• Does not modify official Codex install / asar / signatures
• Does not silently write API Base URL or keys
• CDP remains loopback-oriented (127.0.0.1)

Notes / 补充

• Exact candidate: 88d853b1a9854b4a7b34fcd78f8c81d2b8f8f269
• Fresh independent exact-HEAD review: APPROVE, no P0–P3 findings.
• No rebase was performed. The existing lt_home_panel_compat branch was used as requested, so this PR also contains its two earlier home-layout commits.
• Full test without the skip flag reaches Doctor and reports the intentional source-vs-installed CSS customization difference; the test suite itself is green.